Single-pixel imaging via data-driven and deep image prior dual networks

Area of Science:

• Optics and Photonics
• Computational Imaging
• Machine Learning for Imaging

Background:

• Single-pixel imaging (SPI) reconstructs images using a single-pixel detector, often relying on deep neural networks.
• Deep image prior networks (DIP-Net) offer quality but require many iterations; data-driven networks (DD-Net) are fast but need similar training data.
• Sub-sampling conditions in SPI reduce effective information, challenging image reconstruction quality.

Purpose of the Study:

• To introduce a novel dual-network iterative optimization (SPI-DNIO) framework for enhanced single-pixel imaging.
• To overcome the limitations of existing DIP-Net and DD-Net approaches in SPI.
• To improve image reconstruction quality and reduce iteration count in SPI, especially at low sampling rates.

Main Methods:

• Developed a dual-network iterative optimization (SPI-DNIO) framework integrating DD-Net and DIP-Net strengths.
• Designed a residual block with gradient information to enhance deep network learning from low-sampling-rate SPI data.
• Conducted indoor active lighting and outdoor passive lighting experiments to validate the framework.

Main Results:

• The SPI-DNIO framework achieved high-quality image reconstruction with significantly fewer iterations compared to traditional methods.
• The gradient-enriched residual block improved the network's ability to learn from SPI inputs with less information.
• Experimental results demonstrated exceptional reconstruction capabilities and strong generalization performance across different lighting conditions.

Conclusions:

• The proposed SPI-DNIO framework effectively combines the advantages of DD-Net and DIP-Net for superior SPI performance.
• The enhanced residual block design addresses the challenge of limited information in low sampling rate SPI.
• The framework shows promising results for both active and passive lighting scenarios, highlighting its versatility and effectiveness.